THE ACCOUNTING REVIEW Vol. 65. No. 1 Januaiy 1990 pp.

49-71

The Relation Between Stock Returns and Accounting Earnings Given Alternative Information
Robert Lipe
University of Michigan
ABSTRACT: This paper examines the relation between stock returns and accounting earnings under the assumption that the market observes current-period information other than earnings. This assumption is motivated by existing empirical evidence that stock returns lead accounting earnings. The analysis shows that the returns-earnings relation depends on the relative ability of earnings versus alternative information to predict future earnings as well as the time-series persistence of earnings. Assuming that the researcher does not observe the alternative information, the earnings response coefficient should be increasing both in the ability of past earnings to predict future earnings and in earnings persistence. The variance of stock price changes during the announcement of earnings should be decreasing in predictability and increasing in persistence. Empirical tests of these four hypotheses are generally consistent with the theory. Also discussed is how the assumption of alternative information may be useful in examining the information environment hypothesis, in assessing ad hoc methods of reducing measurement error bias, and in formulating how economic earnings differ from accounting earnings.

HIS study investigates the relation between stock retums and accounting eimings under the assumption that the market observes current-period information other than earnings that is useful in predicting future earnings.

I am grateful to Vic Bernard, Dan Collins. Gene Imhoff. Bill Kinney. Roger Kormendi, S. P. Kotharl. Jerry Lobo, Evelyn Patterson, Bill Ricks, Tom Stober. Sundararaman Thiagarajan. Robert Verrecchia, Jim Wahlen, and Dave Wright for helpful discussions and to the anonymous referees for their Insightful comments. Also, the workshop participants at the University of Michigan, Washington University, and Duke University provided many useful comments. Funding was provided by the University of Michigan and the Peat Marwick Foundation. Manuscript received July 1988. Revisions received February 1989 and June 1989. Accepted July 1989.

49

50

The Accounting Review, January 1990

Beginning with the seminal work by Ball and Brown (1968). the returns-earnings relation has been the focus of many studies. As the literature has progressed, researchers have used finance theories to make empirical predictions. For example. Kormendi and Lipe (1987) model stock returns as a function of the revisions in expectations of earnings, assume that earnings can be represented by a univariate time-series process, and then show that the time-series properties of earnings wiil be an important factor in the returns-earnings relation. Other studies also use explicit theoretical models (Beaver et al. 1980; Beaver et al. 1987; Collins and Kothari 1989; Easton and Zmijewski 1989; Imhoff and Lobo 1988; Lipe 1986). The goal of this area of research is to increase understanding of how earnings and other accounting information are related to stock prices. This paper's focus on alternative information is motivated by recent empirical evidence and theoretical models. Beaver et al. (1987). Collins et al. (1987), and Collins and Kothari (1989). among others, find that imexpected earnings from year t-l-1 (as measured by the researcher) are correlated with returns from year t. The implication is that the market obtains alternative information in year t which is a substitute for some of the "news" in earnings of year t-l-1. Further, Holthausen and Verrecchia (1988) use an information economics model to demonstrate that the relation between stock prices and a given source of information depends on the availability of other useful information. In order to examine the role of information other than accounting earnings. I gissume that during year t the market receives a noisy signal of earnings for year t+1. Combining this assumption with the model used in Kormendi and Lipe (1987). I show that the stock return reaction to earnings is a function of (1) the time-series properties of earnings. (2) the interest rate used to discount expected future earnings, and (3) the relative ability of earnings versus alternative information to predict future earnings. The third factor is the result of allowing the market (but not researchers) to observe alternative information. Four testable hypotheses regarding the relations between stock prices and earnings are derived. The first two are that the coefficient which measures the stock return response to a one-dollar earnings shock (hereafter, the response coefficient) is an increasing function of both the "predictability of the earnings series'* and the time-series persistence of earnings. The predictability of earnings is defined as the ability of past earnings to predict future earnings, and it is refiected in the variance of the shocks in the univariate earnings process (as the variance decreases, the predictability of earnings increases). As earnings predictability increases, the current earnings information becomes more useful in predicting future earnings and. therefore, the response coefficient increases. An alternative interpretation is that as earnings predictability increases, the difference between the univariate earnings shock and the market's issessment of unexpected earnings decreases and. therefore, the downward bias in the response coefficient decreases. The difference between predictability and persistence is that the predictability of earnings is a function of the average absolute magnitude of the annual earnings shocks, whereeis the time-series persistence of earnings refiects the autocorrelation in earnings. The other two hypotheses are that the variance of stock price changes during the release of earnings is negatively related to earnings predictability and

Lipestock Retums and Accounting Earnings

51

positively related to earnings persistence. Holthausen and Verrecchia (1988) and Beaver (1968), among others, analyze the variance of price changes when information is announced. The negative effect of predictability occurs because the variance of price changes measures the retum reaction to the typical earnings shock. As the predictability of earnings increases, the absolute magnitude ofthe earnings shock will probably be smaller, leading to a lower variance of price changes despite the fact that as predictability increases, a given earnings shock is worth more (the response coefficient is larger). The effect of persistence is positive because the greater persistence means a larger reaction to the typical eeimings shock. The fovir hypotheses are empirically tested using the sample firms from Kormendi and Lipe (1987). The parameters of interest (the response coefficient, variance of price changes, and predictability and persistence of earnings) are estimated for each of 145 firms. Simple and partial rank correlations show that the response coefficient is positively related to both predictability and persistence across firms. The variance of price changes is negatively related to predictability, but the positive relation with persistence is very weak. Sensitivity einalyses suggest that the results are not due to ignoring cross-firm differences in risk or size. However, errors in measuring the imivariate earnings shocks may cause overstated significance levels. The rest ofthe paper is organized as follows. The theoretical relation between retums and eimings is developed and discussed in Section I, with the mathematical details presented in the Appendix. Section II presents the empirical tests. Section III discusses other insights gained from a theoretical model that assumes that the market has altemative information. Section IV summarizes and concludes the paper. I. Theory The relations between stock prices, accounting earnings, and altemative information are represented by the following three equations: ..), (1) (2)
(3)

In equation (1), the retum on a share of common stock, R,, is a function of some exogenous expected return, R', and the unexpected retum due to the release of accounting earnings, X,, and the altemative information, I,. Lags of X, and /, are included in equation (1) because they may be useful in determining the imanticipated information. Three assumptions are imposed on equation (1): stock price equals the present value of expected future dividends; the discount rate, |3, is constant over time; and the present value of the revisions in expectations of future dividends equals the present value of the revisions in expectations of future earnings. Together, these assumptions imply that the unexpected stock retum in period t, R',

52

The Accounting Review, January 1990

equals the present value of the revisions in expectations of current and future earnings. The flrst and second assumptions are commonly used in finance and accounting studies. Ohlson (1988) shows that the two represent special cases of a more general model based on state dependent dividends and a no arbitrage equilibrium. The third assumption can be thought of as an extreme version of the statement, "accounting earnings provide information about the future dividend paying ability of the flrm." While this is a stringent assumption, some link between earnings and dividends is necessary in order to derive a relation between returns and earnings. For example, Ohlson links earnings and dividends by assuming that both are driven by the same underlying events or states of nature. The model could be based on less restrictive valuation assumptions, but these three are chosen in order to clearly demonstrate the effect of alternative information.' Equations (2) and (3) specify the information structure of the model. Equation (2) describes the univariate characteristics of earnings. The b, are the autoregressive coefficients, and e, is the serially uncorrelated earnings shock in period t. Kormendi and Lipe (1987) demonstrate that larger bi coefficients cause the current-period earnings shock to have a larger impact on future earnings (greater persistence). The "predictability of the earnings series" is captured by the variance of the earnings shocks, oi. If aJ=O. then past earnings predict fliture earnings perfectly. The ability of past earnings to predict future earnings decreases as ai increasies. Note that the predictability and the persistence of the earnings series are distinct concepts. Persistence describes the time-series relation between the current-period earnings shock and future earnings. Predictability reflects the variation in the earnings shocks. One can imagine a randomwalk and a white-noise series which have the same variance of shocks. The former has much more persistence, but the two are equally predictable (the forecast errors have the same variance). Alternatively, one can imagine two random-walk series with high and low variance of shocks, respectively. The former is less predictable, but the two have equal persistence.* Equation (3) shows that the alternative information equals next period's earnings plus noise. n,+,. The noise is assumed to be serially uncorrelated and is uncorrelated with e,*. for all k. In the spirit of Holthausen and Verrecchia (1988). investors have useful information other than current-period earnings.^ The alter' These are essentially the same assumptions used In Kormendi and Lipe (1987), and using them provides some continuity. In addition, Ohlson (1988, 42) states that, ". .. useful empirical studies can be conceived even when the concepts of what determine security value are unspecified or underidentified, or when the study maintains hypotheses that do not derive from more primitive assumptions." 'The economic determinants of the persistence or the predictability of earnings are not investigated in this paper. Instead, the theory shows that these two parameters are potentially important, and emplrlcaJ evidence suggests that they are important. Given the results, an economic analysis of the peirameters appears useful. For example. Lev (1983) provides some intuition regarding the autocorrelation of earnings and the variability of earnings shocks. ^ Other studies have modeled some forms of alternative information. One approach Is to decompose earnings and show how the components can provide more information than earnings alone (Lipe 1986; Rayburn 1986: Wilson 1986; among others). In most of these models, assessments of expected future earnings do not involve trade-offs between the compwnents and, therefore, earnings predictability is not a factor. Another approach is to include both accounting earnings and

Lipestock Returns sind Accounting Earnings

53

native information could encompass production/investment decisions or analysts' forecasts of X^i made at t. Alternatively, since X, and I. are the only sources of new information, one could think of I, as representing all information other than current and past accounting earnings. Since the empirical tests use of annual data. I, and X, are assumed to be observed simultaneously in year t.* The appendix transforms these assumptions into a model of returns as follows:

S^Z^}^^tEM},

(4)

The term in brackets reflects the new information revealed in yeeir t. The unexpected information regarding the current-period earnings shock equals e,Mw,. In year t, the market observes X, and infers e,. But a portion of e, was anticipated in year t-1. Speciflcally. by observing X^, and 7,-i. the market infers w,=e,+n,. The expectation of e, in year t-1 is Mw,, where: M=-. The other new information concerns next year's earnings. By observing X, and I,, the market infers lu^i The expectation of e^, in year t is Mw,*i. Prior to year t, the expected VEilue of ern is zero. Thus, the new information about e^n equals Mw^i. This is multiplied by the discount rate in order to obtain the present value of the revision in the expectation of X^n. PVR is the persistence of the earnings series. If the univariate earnings series were white noise, then PVR=O and equation (4) would only contain the bracketed term divided by price. But if the time-series coefficients differ from white noise, then the new information in X, and I, will affect the expectations of earnings in periods beyond t+1. In other words, the new information persists into the future, and PVR captures this persistence. P,-i is the stock price at the end of year t-1. Equation (4) presents the theoretical relation between returns and earnings. Before the hypotheses can be developed, some estimation issues must be resolved. If researchers observe /,. then they can use the market's assessment of unexpected current-period earnings. e,-Mw,, as the independent variable in esti-

"permanent" earnings In the model (Beaver et al. 1980; Beaver et al. 1987). The typical assumption is that accounting earnings equals the market's perception of permanent earnings plus noise. In this case, permanent earnings appeair to be known or knowable without reference to accounting earnings eind, therefore, the latter do not play a role in the valuation of securities. In contrast, under the Holthausen and Verrecchia (1988) assumptions and the assumptions discussed above, the market revises expectations of future dividends based on accounting eemiings and alternative information. * As with the valuation assumptions, there are alternative ways to structure equations (2) and (3). For example, /, could be a vector of Information. Also, there could be Interactions between X, and /, by assuming that e, cUid n, are cross-correlated or that /,-i {X,-,) affects X, (/,). If one used a shorter cumulation period for stock returns, then the release of /, and X, would be sequential instead of simultaneous, as discussed in the Appendix. The exact effect of changing the assumptions would depend on what new assumptions were made.

54 mating the reponse coefficient as follows:

The Accounting Review. January 1990

^^}^

(5)

But in order to be consistent with previous papers, I estimate the empirical relation using only returns and earnings data as follows:*

R,=ki+ao+UR,, P.-. =.R' -K1+PVR)(1 - M )

P.-1

"'--^"^Y
.-1

(6)

In equation (6), ao is the response coefficient, R' is the average expected return over time, and UR, consists of n,, n,+,,e^i. and any intertemporal variation in expected returns. Note that UR, and e. are independent. Therefore, UR, represents uncorrelated noise. Using equation (6) instead of equation (5) implicitly assumes that researchers observe R, and e, but not /,. Alternatively, researchers observe noneamings information but choose to ignore it. Thus, the focus of this paper is on how the existence of ilternative information which is not used by researchers affects the observed relation between returns and earnings. The tasks of identifying alternative sources of information and incorporating them into the empirical tests are left to future research. Results in Ou and Penman (1989) suggest that such efforts can be successful. In equation (6), ao=(l-l-PVR)(l-Af). According to the Kormendi and Lipe model, ao=(l-HPVR). The difference is due to including/, in this paper. Note that 0 < M s l which implies that ao<(l-hPVR). Consider the two extreme cases of M=0 and M= 1. First, suppose that /, is not useful in predicting X,+, because ffj= 00 (or /, is all noise). As aJ-oo.M-O, in which case Kormendi and Lipe and equation (6) yield the same theoretical value for ao. This occurs because the difference between the models is the inclusion of/,; but if/, is so noisy that it provides no information, then the models are the same. Second, suppose that al=O, in which case /, predicts X,+i without error. In this case, M = l and ao=O: the return reaction to the earnings shock is zero because current-period earnings are useless in predicting X,+,, given /,. Thus, Kormendi and Lipe's (1987) result is a special case of equation (6). The result that ao is a function of (1 -M) can also be interpreted as an errorsin-variables bias. Since the only explicit information source in Kormendi and Lipe is earnings, e, represents both the earnings shock and the market's assessment of unexpected earnings. In this paper, the latter equals e,-Mw,. Ue,-Mw,
Some researchers estimate the response coeniclent using reverse regression In order to avoid the bias In ao due to measurement errors. However. Section III shows that the reverse regression coefficients are also biased. Therefore, reverse regression Is not used In this paper.

LlpeStock Returns and Accounting Earnings

55

(deflated by price) is the independent variable as in equation (5), then ai-il -I-PVR) which is consistent with the prediction of Kormendi and Lipe. But if the earnings shock is the independent variable as in equation (6), then there is measurement error which equals e.-[e,-Mw,) = Mw,. The well-known errors-in-variables result (Maddala 1977, 292-293) is that:' a = a o ' ( l - M ) = ( l + P V R ) ( l - M ) .

Thus, the response coefficient estimated in equation (6) is a biased estimate of (1+PVR), and (1 -M) represents the bias. This discussion of measurement error is not intended to trivialize the role of the predictability of earnings in the relation between returns and earnings. By including I, in the model, market participants have two competing sources of information to use in forming expectations of future earnings. The usefulness of each signal is determined by its relative ability to predict the future. Thus, R"ia equation (4) is a function of M (and, therefore, predictability) whether researchers use I, to measure market expectations or not. The use or nonuse of I, simply determines whether M appears in the independent variable in equation (5) or in the response coefficient in equation (6). Hypotheses The four hypotheses are now developed. Equation (6) shows that ao is a function of persistence and the relative variability of n, and e,. Differentiating ao with respect to ai demonstrates the relation between the response coefficient and the predictability of the earnings series: aao
dai

(l+PVR)ai

(7)

Since al and a] are positive, the derivative is negative as long as PVR > 1. If esimings are a white noise process, then PVR=O. A negative PVR would mean that earnings are less persistent than white noise which, based on prior empirical results, is unlikely. Thus, the derivative is negative. As the predictability of the earnings series increases (or as al decreases), the response coefficient increases. To understand the positive relation, recall that in period t - 1 , the market's expectation of e, equals:
a]

(e,+n,). [ol+al) Decreasing a] while holding al constant implies that less of e, is anticipated. The decrease in al causes the market to rely less on the prior information and more on X,. In other words, as predictability increases, X, is more useful in assessing expectations of future earnings. Alternatively, the measurement error perspec'Thls assumes that deflating the dollar returns and the dollar earnings metrics by P,-, does not affect the expected value of the coefficients.

Mw,=

56

The Accounting Review, January 1990

tive suggests that ao is positively related to predictability because increasing earnings predictability reduces the variance of the measurement error which causes a decrease in the downward bias in ao. Since 0<M< 1, equation (6) shows that Oo is positively related to PVR except in the extreme case of M = l . An increase in the time-series persistence of earnings leads to an increase in the response coefficient. An alternative method of analyzing the relation between stock prices and earnings is to examine the variance of price changes around earnings announcements. The change in price. AP,, equals the stock return, R,, multiplied by the beginning of period price. P^..' Equation (4) implies that the variance of price changes equals:*
(8)

Note that Var(AP,) is not conditional on the information released in year t. It reflects the average squared change in price based on the mean and variance of the information system and not the price change in response to a given earnings shock. The relation between Var(AP,) and the predictability of the earnings series is assessed by differentiating equation (8) with respect to ai. or:

Since the derivative is positive. Var(AP,) is a decreasing function of the predictability of the earnings series, even though the response coefficient is an increasing function of predictability. Intuitively, if firm A's earnings shocks are less variable than firm B's. a one-dollar shock in each series would cause a much larger stock return for firm A. But on average, the absolute magnitude of the earnings shocks is smaller for firm A and. therefore, the average variation in the price of firm A is smaller. Thus, analyzing stock returns conditional on the fact that earnings are announced during the period is very different from analyzing
'Actually, part of the return In period t may be In the form of a dividend and. therefore. R.P,., may differ from AP.. However, the fonner measure may be more appropriate for addressing the relation between stock prices and earnings; the price changes associated with the ex-dividend date are probably not a function of information released during the year. The empirical tests were also performed using Var(P,,-P,,.,). and the results were similar. ' First, since the variance equals the actual minus expected price change squared. R: does not appear in equation (8). Second, from equation (4). the change in price due to X, and /, is: RrP,.Ml+PVR][e.-Mw,+mw,.,]. The variance of the change in price is then: Var(R,"P,.,] =

=(1-I-PVR)'which yields equation (8).

Llpestock Returns and Accounting Earnings

57

returns conditional on the magnitude of the earnings shock. In particular, these two constructs have opposite relations with the predictability of the earnings series.' A further examination of equation (8) reveals that Var(AP,) is positively related to PVR (again assuming PVR> -1). Thus, whether the relation between stock prices and earnings is assessed using the response coefficient or the veirlance of price changes, the impact of persistence is predicted to be positive. To summarize, the theory predicts that the response coefficient will be positively related to the persistence and the predictability of the earnings series. The unconditional variince of price changes is expected to be positively associated with persistence but negatively associated with predictability. These hypotheses are tested in the next section.

n. Empirical Results
Kormendi and Lipe (1987) estimate the relation between returns and earnings and test their proposition that ao should be related to PVR. This section uses their sample of firms to test whether Oo Is also related to the predictability of earnings. First, the Kormendi and Lipe data and results are discussed. Second, the impact of the predictability of earnings is examined. Third, some sensitivity analyses U"e conducted. Kormendi and Lipe use a two-equation system to estimate both the response coefficient, Oo, and the time-series properties of earnings, b,. Using the notation from Section I, they estimate the following: ^ (10)

(11) The inclusion of the J subscripts indicates that a separate system is estimated for each flrm,J. Also, the autoregressive model of earnings is limited to two lags. In order to reduce the cross-sectional correlation in the data, Kormendi and Lipe use flrm-speciflc returns and earnings. Rj. are the residuals from an annual market model in which the percentage return for firm J*s common stock in year t is regressed on the percentage change in the CRSP value-weighted market index. The annual returns are cumulated from April of year t until March of year t + 1 ,

'The negative relation between the predictability of earnings and Var( AP,) may appear eontrary to the Holthausen and Verrecchia (1988) information economies model. They derive a positive relation between the ability of an information souree to prediet the future liquidating dividend of the firm and Var( AP,). The difference in signs is due to a difference in assumptions. In their model, the firm's value is a function of the liquidating dividend, u. If the ability of earnings to predict u Increases, then the announcement of earnings provides more information. The present assumptions imply that the revisions in expectations of future accounting earnings are valued by the stock market. As mentioned above, as the predictability of the earnings series increases, the release of X, cind /, contains less new Information regarding future earnings and. therefore, Var(AP,) decreases. Indeed. Holthausen and Verreeehia show that decreasing the variance of u leads to a decrease in Var(AP,). Thus, the two models agree that as the variance of the attribute that is valued in the stock market decreases, Var{AP,) decreases,

58

The Accounting Review, January 1990 Table 1 Summary Statistics for Estimated Parameters' (n= 145 Firms)

Model: (10)
AXy, =
it-i+ej.

(11) Third Quartile 5.07 10.12

.68

Standard
Parameter
PVRj
^1

Mean 3.38 8.93

.78

Deviation

Minimum

First Qvuxrtile 1.41 6.92

.06

Median 2.50 7.91

.17

Maxtinum

bu

i>v
\j

MEDVALJ

45.58 -.08 -.17 1.10 1239

3.12 3.41 1.84 86.69

.29 .24 .34

-2.28 4.36
.00

2.33 -.70 -.79

.32 9.3

13.20 -.29 -.31

.87 115

23.11 -.07 -.19 1.09

397

43.32
.12

17.98 22.71 17.83 774.4

.92 .66

-.06 1.28
772

3314

2.09 27695

Ry,=the flrm specific (real) percentage return for flrm J In year t. AXy,=the flrm speclflc change in (real) earnings. P,,-,=the (real) stock price for flrmj at the beginning of year t. ey,=the shock In the univeiriate earnings series. bu=the autoregressive coefficients. a(y=the response coefficient. Other estimated parameters sire as follows: i*W?y=eamings persistence, derived from the estimated by, assuming T= 10%. ^^=variance ofthe estimated earnings shocks, 6j,. 9iPj=vai\ance of Rj,xPj,.,. Xy=estimated beta from an annual meirket model regression. MEDVALy=the median value of equity for each flrm.J, in millions.

'

because sample firms all have December 31 year-ends.'P/,-i Is the stock price at the beginning of the year t cumulation period. Similarly, AXj, are the residuals from a regression of dollar changes in earnings per sheire (before extraordinary items) for flrmJ on the changes in the Standard and Poor's index ofearnings. Rj,, Pjt-i, emd AXy, are adjusted by the consumer's price index in order to mitigate the heteroscedasticity caused by inflation (see Kormendi eind Lipe 1987 for a more complete description ofthe data). Data from 1947-1980 are used to estimate the coefficients of the system (lO)-(ll) for each of 145 firms. A nonlinear weighted least squares approach is employed. As mentioned above, the existence of alternative information implies a o ^ d +PVR). Table 1 contains summary statistics for the estimated parameters.
' The April-March cumulation has been used in prior studies so that the returns are contemporaneous with the three qucirterly and one annual earnings announcements for the yeeir. The correlations between earnings and returns are larger for cumulation periods which begin in year t - 1 and last longer than 12 months (Collins et al. 1987). However, since the purpose of this paper is to examine the contemporaneous relation between R, and e, under the assumption that the market has alternative information, the April-March cumulation is appropriate.

Lipestock Returns and Accounting Earnings Table 2 Kendall r Rank Correlations (n= 145 Firms)
PVRj

59

Pvki

%-Rj

PVR]
MEDVALJ

.211" .018 -.229" .285" .109'

-.256" .286''* .083 -.211" -.083

.005* -.157' .584" .025

-.001 .015 .209"

-.573" -.041

.046

(t,=the estimated response coefilcient for flrmj. y = the time-series persistence of earnings forflrmJ assuming T= 10%. j =the persistence of earnings based onflrm-specificinterest rates. ^(/=variance of the estimated earnings shocks, 6j,.
(^4Pj=variance ofRj,xPj,.i.

\y=estimated beta from an annual market model regression. MEDVALJ=the median value of equity for eachfirm,j , in millions. ' Signiflcant at the .001 level. ' Signiflcant at the .01 level. ' Significant at the .10 level. * Signifies the four direct tests of the model. These four significance levels are based on one-tailed tests. All other significance levels are based on two-tailed tests.

The mean (median) of doy across the 145 firms Is 3.38 (2.50). The mean (median) of PVRJ is 8.93 (7.91). In addition, agyjs less than {1+PVRj) for 144 of the 145 firms. Note that the calculation of PVRj requires an assumed interest rate for discounting expectations of future earnings; a rate often percent is used in Table 1. The results are consistent with equation (6).'* Tests of Earnings Predictability The response coefficient, ao, is hypothesized to be an increasing function of the predictability of the earnings series. Predictability is reflected in the variance of the ecimings shocks, ai (as ai increases, the predictability decreases). The empirical measure of ai for firmj is denoted ^,j, and it equils Var(ey,), where &j, are the estimated residuals from equation (11). The estimates CLOJ and ^.j can be viewed as random variables, the distribution of which depends on the "true" distribution of aoj and aij, respectively, and the errors from estimating the parameters for each firm. These estimates are analyzed across firms in order to infer the relation between the "true" aoj and aij. In testing the hypothesis, the null is that CLoj and V,j are unrelated (or positively related), with the alternative hjrpothesis that the two are negatively related.
" since PVR ls a decreasing function of the interest rate, assumlng^an Interest rate substantially higher than ten percent will result In Aq, being greater than (l + PVRj) for most firms. However. Kormendi and Lipe (1987,341) report that even If one assumes a rate of 30 percent, CiQi<(l+PVRj) for 67 percent of the firms.

60

The Accounting Review. January 1990 Table 3 Kendall Partial Rank Correlations (n= 145 Firms)

-.335" .297'

.130' .084'

6Q,=the estimated resjwnse coefficient for flrmJ. f*VRy=the time-series persistence of earnings for flrmj, assuming T= 10%. O^=varlance ofthe estimated earnings shocks. &j,. OA,J=variance of Ry,xP,,.,. ' Significant at the .001 level (one-tailed test). ' Significant at the .02 level (one-tailed test). ' Significant at the .10 level (one-tailed test).

Because the functional form of the relation between OQ, and i^,j is nonlinear, the relation is examined using rank correlations." Table 2 presjents the Kendall T rank correlations for the data. The correlation between doj and V.j is - .371 with a Z-statistic of - 6 . 6 . The correlation is significantly negative at less than the .001 level." The null is rejected in favor of the alternative. Table 2 shows that the correlation between aoj and (1 +PVRj) is significantly positive, which is the same result reported by Kormendi and Lipe.'* But note that i^.j and [1+PVRj) are significantly negatively correlated (rank correlation = -.256, Z= -4.57). Recall that PVR is a function of the autoregressive coefficients ofthe earnings series. The negative correlation is apparently showing that firms with less persistent time-series have larger variances of earnings shocks. The theory in Section I makes no prediction regarding the sign or magnitude of the relation between i^.j and {1+PVRj), but the negative relation is consistent with Lev's (1983) analysis of economic factors such as whether the firm produces durables or nondurables^ Since ^.j and [1+PVRj) are significantly correlated, the simple correlations between CLOJ and these two variables may suffer from an omitted variables bias. In other words, the simple correlation between, say, aoj and (1 +PVRj) could be overstated because movements in (1+PVRy) are proxying for movements in ^,j (Maddala 1977, 155-157). Therefore, partial rank correlations are presented in Table 3 in order to assess the incremental importance of persistence and predictability. The partial rank correlation between doj and ^.j, given (1+PVRj), is
" Scatter plots (not reported) show a strong curve in the relation as well as some outliers. The rank correlations are less sensitive to these problems than Pearson correlations or OLS regressions. "Under the null, Z~N{O,1). While using rank correlations reduces the impact of outliers and nonllnearitles, the test does assume that each pair, d^, and P^. represent an independent draw. Crosssectional correlations among the firms could result in the estimated parameters being crosssectionally dependent, however, the market-wide movements were removed from earnings and retums to reduce this dependency. " The table reports the^rrelations between various peirameters andf'WJy, which are the seime as the correlations with H+PVRj).

Lipestock Returns and Accounting Earnings

61

-.335. with Z= -5.93 (significant at .001);" again the null is rejected which is consistent with the hypothesis developed in Section I. The correlation between O y and (1+PVRy), given i^.j, is .130. with Z=2.30. which is significant at .02 o using a one-tailed test. Thus, both predictability and persistence appear to be significant determinants of the response coefficients. Section I also analyzed the variance of price changes. The theory shows that Var( AP.) should be positively related to both aJ and (1+PVR). Let i^^pj represent the variance of Rj,xPj,.t, estimated over time for each firm. Table 2 shows a simple rank correlation between t^^py and 9^ of .286. Table^3 shows that the partial rank correlation between i^^pj and ^^. given (1+PVRj), is .297. Both correlations are significant at .001. Thus, the empirical evidence supports the hypothesis of a negative relation between the variance of price changes and the predictability of the earnings series. The simple rank correlation between (1 +PVRj) and P^PJ IS essentially zero. The partial correlation between these two variables, given \^.j, is .084 which is significantly different from zero at .07. Persistence is not as strongly related to the variance of price changes as it is to the response coefficient. Sensittvity Analysis The empirical tests are Joint tests of the null hypotheses and the assumptions. If the assumptions are poor representations of the interactions between accounting earnings, stock prices, and alternative information, then equation (6) will be a poor description of ao. The major concern is whether the results are spurious. Note, however, that deviations between the model's assumptions and the real world would likely result in a failure to reject the null instead of a spurious rejection. An exception to this general rule is if the assumption involves an omitted correlated variable, as discussed below. Also, because this study uses annual data, it can only test whether the association between returns and earnings are consistent with the hjrpotheses; the tests cannot determine whether earnings cause returns. Measurement errors have a subtle effect on the tests. As discussed in Section I. &0J is a downwardly bisised estimate of (1 +PVRj) because one cannot observe the noneamings infonnation used by the market. But the inability to measure market expectations does not bias OQ, away from its theoretical value of (1 -I-PVR>)(1 -Mj). On the contrary, the hypothesized positive relation between a<v and 9.J explicitly incorporates the informational superiority of the market. Therefore, the traditional errors-in-variables bias mentioned in most response coefficient studies (Kormendi and Lipe 1987. 331) can not lead to spurious results in this paper. The estimate of aoy will be affected by errors in specif3ring the univariate timeseries model ofearnings. Suppose the estimated earnings shock. e>,. contains the true shock, ej,, plus measurement error, ej,. The efiiect of ej, is to bias the Oo^ towards zero, and the^ias is an Increasing function of al. But the measurement error also means that V^=oij+al. If there is substantial cross-firm variation in al.
" Maghsoodloo and Pallos (1981) describe hypothesis testing with partial rank correlation. The normal approximations used here are Interpolated which might slightly understate the Z-statistic.

62

The Accounting Review, January 1990

then firms with larger al will have larger P.j but smaller ap,. Therefore, the negative correlation between cioj and i^.j could be due to measurement errors in the earnings shocks instead of due to a negative relation between ao and al. Unfortunately, the true earnings shocks and their variance are not observable and, thus, direct resolution of this issue is impossible. However, the restilts regarding the variance of price changes can provide indirect evidence. Note that PAPJ is not a function of al because the former is not conditional on the estimated earnings shocks. If a large portion of the cross-sectional variation in V,j is due to al, then the correlation between i^^pj and Vej would be indistinguishable from zero. The^mpirical results presented above show a sigificant correlation between 9APJ and V.J, which suggests that \^.j captures significant cross-firm variation in al. This suggestsjbut does not guarantee) that the significant negative relation between cioj and V,j is not spurious. One assumption that could lead to spurious correlation is the cross-sectionally constant interest rate assumption. Differences in interest rates would refiect differences in risk. Omitting finn-specific risk affects the tests in two ways. First, if r varies across firms, then PVRj as calculated above is incorrect. To assess the severity of this error, the estimated market model beta of firm J, denoted %, is used to calculate a fimi^pecific interest rate, fy." Table 2jhows that the new persistence measure, PVRJ, and the original measijre, PVRj, are significantly correlated. The simple rank correlations between PVRJ and both do> and V^pj actually increase, and the partial correlation between PVRj and aoj, given V.^, is more significant than before. Thus, incorporating firm-specific risk into persistence appears to provide stronger results. The second effect of differential risk is that if my valuation assumptions are not valid, ao may be related to risk in ways other than in equation (4). Indeed, Collins and Kothari (1989) find that the estimated response coefficient is negatively related to the firm's stock market beta. Table 2 shows that the rank correlation between dq/ and %j is -.229 (significant^t the .001 level). However, the partial correlation between do^ and Xj, given PVRJ, is only -.084 (significant at the .07 level, using a one-tailed test). Thus, once PVR is adjusted for differences in Xy, the incremental explanatory power of X is greatly reduced. Another concern is y that V.y could be proxying for %. But this is not the case as the rank correlation between V.j and % is .083 which is not significantly different from zer(a at the . 10 level (two-tailed test). In addition, the partial correlations of doy and PVRj, doy and V,j, and Vipy and V.y, given \j, are all significant in the predicted direction. Thus, the primary results of this paper are probably not driven by omitting cross-sectional differences in risk. Another potential omitted factor is firm size. The contemporaneous relation
" The cedculation of fy Is as follows:

145 Thus, the_ average fj is still ten percent. But now, 90 percent of each firm's Interest rate is determined based on %, and ten percent is a constant. Using factors other than .01 and .09 yields similar results.

Lipestock Returns and Accounting Earnings

63

between returns and earnings has been shown to differ with size (Collins et al. 1987; Freeman 1987). The firms are ranked according to their median market value of equity over the 1947-1980 period, denoted MEDVALj. Table 2 shows that the correlation between MEDVALj and both ctoy and V^^^^are significantly positive. However, size is not significantly correlated with PVRj or V.j- Partial rank correlations (not reported) confirm that the inferences drawn above regarding the effects of persistence and predictability do not appear to be driven by size.'^ These results regarding size may not generalize to the population of firms, however, because this sample contains NYSE firms that survived for at least 34 years.

m. Extensions
Previous sections developed and tested specific hypotheses regarding the impact of alternative information. This section examines some ofthe other benefits of assuming that the market has information other than earnings. First, a model which includes alternative information can address the information environment hypothesis and other issues in a more formal way. Second, since the differences in the infonnation sets observed by the market and the researcher are explicit in the model, one can assess whether ad hoc ways of dealing with measurement errors will be useful. Third, while the assumed link between future earnings and future dividends is without error, the existence of alternative infonnation suggests that economic earnings differ from accounting earnings, and the difference is negatively autocorrelated. Adding alternative information to an existing model of returns implies that the relation between returns and earnings is a function ofthe predictability ofthe earnings series. But equation (4) also shows that the relation depends on al, the ability of the alternative infonnation to predict future earnings. For example, an increase in ai means that J, is not as useful in predicting e^, and, therefore, ao increases. The negative relation between the predictive ability of alternative information and the response coefficient is similar to the information environment hypothesis (e.g., Atiase 1985, 1987; Collins et al. 1987; Collins and Kothari 1989; Freeman 1987; McNichols and Manegold 1983). The difference is that the latter is usually stated in terms of the quantity of alternative information, whereas the former is based on the predictive ability or "quality" ofthe alternative information. Thus, including proxies for how well future earnings can be predicted by alternative infonnation may be useful in testing the infonnation environment hypothesis. Also, one could test whether cross-sectional differences in size, the predominant proxy for differences in infonnation environments, becomes insignificant when cross-sectional differences in ai are controlled. Note that prior studies of the infonnation environment hypothesis generally do not derive their tests from a theoretical model of the relation between returns and earnings, whereas the importance of al is a direct implication from equation (4).
" The results of the four hypotheses tests are similar If mean, beglnnlng-of-perlod, end-of-period. or mlddle-of-period value Is used instead of the median. However, the correlation between dq, and size is not significantly positive for all definitions.

64

The Accounting Review, January 1990

As to measurement error, ao is a function of (1 -M) because the researcher does not perfectly measure the information used by the market. One way of expressing this is that ao is a biased estimate of the true parameter of interest, (l-l-PVR). But note that reverse regression will also yield bieised estimates of (1 -I-PVR). To see this, rewrite equation (4) as:

=kx+k.R'.+k.[l+PVR) ^'M^'^^M^'^^ ^ue,,

P.-1

(12)

Ue, is the residual, andfciand ki are the reverse regression coefflcients. Assuming that Rf is a constant and that scaling doUeir retums and dollar earnings shocks by P,-i does not affect the expected veilue offc2,'*the value of the reverse regression estimate of ao,ao, can be determined as follows: ,13) Thus, ah is a biased estimate of (1+PVR), and the bias depends on M. Since 1-M the bias is away from zero. The bias occurs because R, includes PMw^i which is the revision in expectations of erf i. Since i8Mu;,+t is uncorrelated with e,, a portion ofthe independent variable in equation (12) is uncorrelated with the dependent variable leading to the classic errors-in-variables bias in IC2. Therefore, even if the earnings shocks and stock returns are measured without error, the reverse regression produces biased estimates of (1 -I-PVR) because stock returns reflect information about next year's earnings shock." Again, this becomes apparent when the market is assumed to observe altemative information. Brown et al. (1987) suggest an alternative control for measurement errors using retums prior to the cumulation period for R, as a second independent variable in equation (6). This procedure should reduce the errors-in-variables bias if prior returns are correlated with the measurement error in the proxy for unexpected earnings but are uncorrelated with current-period retums or the market's unexpected earnings. Here the difference between e, and the market's assessment of unexpected earnings equals MUJ,. Equation (4) shows that R,., will be correlated with Mio,. Further,R,-i is uncorrelated withe,Mu;,andR,. Therefore, the model in this paper suggests that the Brown et al. method should be useful.^"
" These assumptions simplify the analysis. If they do not hold, the bias in kz will be more complex, but still exists. " Beaver et al. (1987. 150-151) state that the reverse regression will not provide unbiased coefficients if stock price changes reflect information other than current-period earnings. This analysis shows that such a bias is to be expected. ^ Using quarterly data and a maximum return cumulation period of 40 trading days. Brown et al. (1987) report modest improvements from including lagged returns.

Lipestock Retums and Accounting Earnings

65

Equation (4) also has implications for studies which analyze the differences between economic earnings and accounting ecirnings (Beaver et al. 1980; Beaver et il. 1987).^' In these models, stock price is a function of economic earnings (which is unobservable to the researcher) and accounting eeimings equals economic earnings for the period plus noise. This paper does not explicitly address the noise in accounting earnings because I assume a direct link between the revisions in expectations of dividends and earnings. However, the new information provided by accounting earnings alone, e,, is not equal to economic earnings. The difference between the economic earnings of year t and e, is: {e,-Mw.+0Mw^i)-e,=PMw^i-Mw,. (14) Equation (14) demonstrates that economic earnings lead accounting earnings. Further, the difference is negatively autocorrelated. The reason is that the income from a given transaction will be recognized by both accounting earnings and economic earnings, but the recognition can take place in different years. For example, if economic earnings recognize a portion of a transaction's income in year t and the remaining portion in year t+\ while accounting earnings recognize all ofthe income in period t+1, the difference between the two must be negatively autocorrelated.^^ Thus, by assuming that the market has alternative information, one difference between economic and accounting earnings becomes clear. IV. Conclusions This study examines the theoretical relation between stock returns and accounting earnings, assuming that the market has a second source of currentperiod information in addition to earnings. Theoretically, the stock return during the period is a function of (1) the time-series persistence ofthe earnings series, (2) the interest rate used in discounting expected future earnings, and (3) the relative ability of earnings versus alternative information to predict future earnings. Comparative statistics show that the response coefficient is an increasing function of the ability of past earnings to predict future earnings and an increasing function of persistence. In addition, the variance of stock price changes conditional on earnings being announced is a decreasing function of the predictability of the earnings series and an increasing function of earnings persistence. The predictability/response-coefficient effect is positive because the value attached to a one-dollar current-period earnings shock is an increasing function of predictability. The predictability/variance-of-price-changes effect is negative because the average quantity of unexpected information released during the period is a decreasing function of predictability. The sample from Kormendi and Lipe (1987) was used for direct tests of these four hypotheses, with three being supported by the data. The exception is that
" Some studies refer to economic earnings as "permanent" earnings. Whatever the name, the construct is that number which is valuation sufficient. Also, the difference between economic earnings and accounting earnings is usually termed "garbling." "See Beaver (1970) for a more detailed discussion of how differences between accounting and economic earnings will be serially correlated.

66

The Accounting Review, January 1990

the variance of price changes is only weakly related to earnings persistence. Sensitivity analysis suggests that the results are not driven by omitting risk and size from the empirical tests. Also, by explicitly including alternative information in formulating the hypotheses, the results cannot be attributed to errors in measuring market expectations. But other econometric problems cannot be resolved and, therefore, the strength of the results may be overstated. The model presented and tested in this paper extends the results from previous work. First, existing empirical evidence that earnings shocks in year t+1 are correlated with stock returns in year t suggests modelling the alternative information as a noisy signal of future earnings. Second, previous models and the associated empirical tests demonstrate the importance of earnings persistence. Third, models of how rational individuals assess expectations based on competing sources of information suggest that predictive ability determines the extent to which a given signal is used. Combining these elements provides new insights into the relation between stock prices and accounting earnings. Appendix This appendix contains detailed derivations of the theoretical model. Stock price is assumed to equal the present value of expected future dividends, or:

In equation (A. 1), P, is the stock price per share in period t, E, represents the expectations conditional on all information available at time t, D,*, is the dividend paid to equity holders in period t+s, and /3= 1/(1 -I-T), where T is the appropriate rate of interest for discounting expected future dividends. Dividends are used in equation (A. 1) because, as Ohlson (1988,29) points out, "ultimately only payoffs count, and dividends alone can be consumed." The discotmt rate is asstuned constant across time for simplicity. Total stock retvirns can be decomposed into expected and unexpected components as follows:

J\,
Pfl

Pr-l

This analysis focuses on i?,", the portion of the return in period t that is unexpected based on the information available in period t - l . Combining equations (A.1) and (A.2) yields:
s=0 Pt-i

In order to bring accounting earnings directly into equation (A.3), the present

Lipestock Returns and Accounting Earnings

67

value of the revisions in expected dividends are assumed to equal the present value of the revisions in expected accounting earnings, in which case: (A.4) Obviously the assumption will not hold exactly and, thus, there is some error introduced in equation (A.4). These errors will affect the empirical tests, but without knowing exactly how the present value ofthe revisions in expectations of accounting earnings and dividends differ, it is impossible to accurately predict the magnitude or even the direction of the effect. Equation (A.4) determines the magnitude of the return conditional on the new information released during period t. Equations (2) and (3) in the text describe the information structure. The earnings series is assumed to follow a first-differenced, finite-order autoregressive process. The autoregressive coefficients, bi, axe assumed to be known by both the market and the researcher. The shock (or innovation) in the earnings series, e,, is assumed to be unconditionally distributed JV(O,o^). The normality assumption is convenient later in calculating conditional expectations. The alternative information provides a noisy signal of future earnings. The noise, n,+,, is assumed to be distributed N(0,o^). Also, e,+, and n,+ are assumed to be independent, for all I and k. I, and X, are observed simultaneously because tests in the paper use annuEil data. To understand the contemporaneous relation between returns and earnings, the revisions in the expections of X,*, in equation (A.4) must be restated. It is helpful to rewrite X, in its infinite-order moving-average form, as follows: X.=e{L)e., (A.5) wherefl(L)= 1 +diL+e2L^+. . ., and L is the lag operator. The d coefficients come from inverting equation (2) into a moving-average process. The relation between dj and bi will be demonstrated later. Using equation (A.5), the revision in the expectation ofX,^ is:

= g ejlE.{e.^.j) -E..rle.^j)].

(A.6)

Equation (A.6) can be simplified by determining what new information is released in period t. Note that all past realizations (X,-, and /,.(, 1=1 oo) are known prior to period t. Thus, the release of X, implies that the value of e, is revealed. Further, combining equations (2) and (3) yields:
N

I,=X,n+n^i=X,+J^ b.AX,-,-i-e,+,+a,+,.
(=1

(A.7)

Since the only unknowns on the right-hand side of equation (A.7) are e,+, and n,*,, the release of X, and /, together implies that the sum (e,+,--n,) is also revealed. Note that investors do not know e,+, or n,+, individually in period t. For notational convenience, let u;,t,=e,+,-l-n,. Expanding the right-hand side of equation (A.6)

68 shows that:

The Accounting Review, January 1990

s-2 J=0

(A.8) The summation overj=0 to s - 2 equals zero because e, and iu,+, provide no new information about the earnings shocks in period t-i-2 or thereafter. The summation over /c= 1 to 00 also equals zero because earnings shocks in or before period t - 1 are already known. Thus, only the two middle terms will be nonzero, which yields: +&^i[E.(em)-E^,(e,)]. (A.9) Disclosing e, and u>,+, causes the market to revise its expectations of e, and e^i. Equation (A.9) shows that the revision in the expectation of e,[e^i) leads to a revision in the expectation of X,^, and the magnitude ofthe latter revision is determined by d.(6.-i). Substituting equation (A.9) into equation (A.4) and rearranging terms shows that:

Equation (A. 10) can be simplified in two ways. First, the discounted sum of the moving average coefficients captures how current information impacts expectations of future earnings. This sum is referred to as the time-series persistence of earnings and is denoted PVR, in order to be consistent with the notation in Kormendi and Lipe (1987). They demonstrate that this sum is a function of the b, coefficients, as follows (pp. 329-330):

Second, under the assumption that e, and n, are distributed N{O,ai) and N{O,ai), respectively, E,-,(e,) has the following convenient representation: E...(e,)=^UJ,. ai+ai (A. 12)

LipeStock Returns and Accounting Earnings

69

To prove this, recall that the unconditional expectation of e, equals zero. But at time tl, the market knows w,=e,-i-n,. Since e, suid n, are independent and normally distributed, w, is also normally distributed. The expectation of e, conditional on w, is then:

In equation (A.13), n, In,,) equeds the unconditional expectation of e, [w,); a. (a.) equals the standard deviation of e, {w,); and p equals the correlation between e, and w,. Since and p=equation (A. 12) holds. For convenience, let

Using equation (A. 12), the revisions in expectations of e, and e,*, are as follows:

These two simplifications transform equation (A. 10) as follows:

Equation (A. 15) represents how new information impacts stock returns. The relation between returns, earnings, and alternative information in equation (A. 15) presumes that X, and J, are observed simultaneously. While this is appropriate when annual returns are used, the model could be specified differently. For example, if returns are cumulated over a few days surrounding the earnings announcement, then one would naturally assume that X, is released during the cumulation period. However, the relation in equation (A. 15) would hold for the short cumulation period only if the alternative iniformation is released during the cumulation period and no information is released between cumulation periods. A more reasonable assumption is that /, is revealed during the time between cumulation periods and X, is released during the cumulation period. This leads to a sequential release of information that is similar to the sequence considered by Holthausen and Verrecchia (1988). If information is released sequentially, then the form of equation (A. 15) changes, and the change depends on the alternative information assumption (eq. [3]). For example, one could replace equation (3) with: I,=X.+n.. (A. 16) In this case, /, reveals e,+n, at the beginning of period t instead of revealing e,*i+n,*i. The relation between returns cumulated over the short period of time and e, will still be a fimction of M, so the predictability effect is still present. But

70

The Accounting Review, January 1990

the correlation between R, and e,+i Is zero. Thus, the lagged return effect that motivates the form of equation (3) is eliminated by assuming equation (A.16). If one maintains equation (3) and Edso assumes sequential information, then the model becomes much more cumbersome. At the beginning of period t, I, reveals information about e, and e,+,. At the end of the period, X, resolves all uncertainty regarding e, and provides additional information about e,+i. Thus, there are multiple sources of information regarding e,+i which complicates the conditioned expectations. In general, the relation between returns and earnings is still determined by both predictability and persistence. Also, R, is still correlated with References Atiase, R. K. 1985. Predisclosure information, firm capitalization and security price behavior around earnings announcements. Joumai of Accounting Research. (Spring): 21-36. . 1987. Market implications of predisclosure infonnation: Size and exchange effects. Joumai of Accounting Research. (Spring): 168-176. Ball, R., and P. Brown. 1968. An empirical evaluation of accounting income numbers. Joumai of Accounting Research. (Autumn): 159-178. Beaver, W. H. 1968. The infonnation content of annual earnings announcements. Joumai of Accounting Research. (Supplement): 67-92. . 1970. The time series behavior of earnings. Joumai of Accounting Research. (Supplement): 62-99. R. A. Lambert, and D. Morse. 1980. The infonnation content of security prices. Joumai of Accounting and Economics. (March): 3-28. , , and S. G. Ryan. 1987. The information content of security prices: A second look. Joumai of Accounting and Economics. (Juiy): 139-157. Brown, L. D., R. L. Hagerman, P. A. Griffin, and M. E. Zmijewski. 1987. An evaluation of alternative proxies for the market's assessment of unexpected earnings. Joumai of Accounting and Economics. (July): 159-193. Collins, D. W., S. P. Kothari, and J. D. Raybum. 1987. Firm size and the infonnation content of prices with respect to earnings. Joumai of Accounting and Ek:onomlcs. (Juiy): 111-138. , and . 1989. An analysis of intertemporal and cross-sectional determinants of earnings response coefficients. Joumai of Accounting and Economics. (July): 143-181. Easton, P. D., and M. E. Zmijewski. 1989. Cross-sectional variation in the stock market response to accounting earnings announcements. Joumai of Accounting and Ek:onomtcs. (July): 117-141. Freeman, R. N. 1987. The association between accounting earnings and security returns for large and smali firms. Joumai of Accounting and Economics. (July): 195-228. Holthausen, R. H., and R. E. Venecchia. 1988. The effect of sequentiai infonnation releases on the variance of price changes in an intertemporal multi-asset market. Joumai of Accounting Research. (Spring): 82-106. Imhoff, E., and G. Lobo. 1988. The impact of earnings uncertainty on stock price variability and the information content of unexpected annual earnings. Working paper. University of Michigan. (November). Kormendi, R., and R. C. Lipe. 1987. Earnings innovations, earnings persistence, and stock returns. Joumai of Business. (July): 323-345. Lev, B. 1983. Some economic determinants of time-series properties of earnings. Joumai of Accounting and Economics. (April): 31-48. Lipe, R. C. 1986. The information contained in the components of earnings. Joumai of Accounting Research. (Supplement): 37-64.

Lipestock Returns and Accounting Earnings

71

Maddala. G. S. 1977. Econometrics. McGraw-Hill. Maghsoodloo, S., and L. L. Pallos. 1981. Asymptotic behavior of Kendall's partial rank correlation coefficient and additional quantile estimates. Joumai of Statistical Computer Simulation. (41-48). McNichols, M., and J. G. Manegold. 1983. The effect ofthe infonnation environment on the relationship between financial disclosure and security price variability. Joumai of Accounting and Economics. (April): 49-74. Ohlson, J. A. 1988. A synthesis of security valuation theory and the role of dividends, cash flows, and earnings. Working paper, Columbia University. (September). Ou, J., and S. Penman. 1989. Financial statement analysis and the prediction of stock returns. Joumai of Accounting and Economics. (Forthcoming). Raybum, J. 1986. The association of operating cash flow and accruals with security returns. Joumai of Accounting Research. (Supplement): 112-133. Wilson, G. P. 1986. The relative information content of accruals and cashflows:Combined evidence at the earnings announcement and annual report release date. Joumai of Accounting Research. (Supplement): 165-200.